Microsoft used Build 2026 materials to describe MAI-Thinking-1, a reasoning model with 35 billion active parameters, a 256K context window and private-preview availability on Foundry. Microsoft also says the model matches Opus 4.6 on SWE Bench Pro.
Microsoft said in its Build 2026 blog post, “Microsoft Build 2026: Be yourself at work,” that its new MAI-Thinking-1 reasoning model has 35 billion active parameters and a 256K context window.
The company positioned the model around lower token costs, according to the same Microsoft blog post, and said MAI-Thinking-1 is available in private preview on Foundry. Microsoft also claimed in the Build 2026 post that the model’s coding performance matches Opus 4.6 on SWE Bench Pro.
The core technical details Microsoft disclosed are concise but significant for developers evaluating model tradeoffs. In Microsoft’s description, MAI-Thinking-1 is a reasoning model with 35 billion active parameters, which indicates the amount of model capacity active during inference rather than a broader total-parameter figure. Microsoft’s stated 256K context window points to support for long inputs, such as extended codebases, documents or multi-step work sessions, though the company’s excerpted announcement does not provide additional benchmark details beyond the SWE Bench Pro comparison.
Microsoft’s Build 2026 news hub says the company’s official keynote coverage and related AI announcements include Web IQ, Foundry IQ, Work IQ APIs, Frontier Tuning and MAI models. That places MAI-Thinking-1 within a broader Build 2026 slate focused on AI development tools and model access rather than as a standalone release.
Microsoft said MAI-Thinking-1 is available in private preview on Foundry, according to its Build 2026 blog post. The Microsoft AI keynote transcript for Build 2026 also discusses MAI model launches and partner availability, including Baseten and OpenRouter integrations.
Those details suggest Microsoft is presenting the MAI models through both its own developer platform and selected partner channels. However, the provided Microsoft materials do not specify public general-availability timing, pricing, exact usage limits or the full methodology behind the SWE Bench Pro comparison.
For developers, the most concrete claims in Microsoft’s Build 2026 materials are the model size, long context window, private-preview status and Microsoft’s coding-performance comparison. A 256K context window can be relevant for workflows involving long source files, large documentation sets or multi-file reasoning. Microsoft’s low-token-cost positioning, if reflected in final pricing, could matter for teams running reasoning-heavy workloads at scale.
Still, the available Microsoft sources leave important evaluation questions open. Developers will likely need access to the private preview, detailed pricing and independent testing before comparing MAI-Thinking-1 with other reasoning models in production settings.
Microsoft’s Build 2026 announcement therefore makes MAI-Thinking-1 a notable addition to the company’s AI model lineup, but the public record so far is limited to Microsoft’s own stated specifications, availability notes and benchmark positioning.
The company positioned the model around lower token costs, according to the same Microsoft blog post, and said MAI Thinking 1 is available in private preview on Foundry.
Microsoft also claimed in the Build 2026 post that the model’s coding performance matches Opus 4.6 on SWE Bench Pro.
What Microsoft is emphasizing The core technical details Microsoft disclosed are concise but significant for developers evaluating model tradeoffs.
Continue reading
